Evaluation of epileptic preictal activities by EEG-Aura based convolutional neural networks and topographic mapping approach

In this thesis study, inspired by the epileptic auras defined in the medical, the reflections of pre-seizure warning symptoms on the EEG signal were evaluated with deep convolutional neural networks (ESA) and the EEG-Aura approach was proposed for the preictal period. The evaluation process was carried out in three stages. These stages consist of detection of the patient-specific EEG-Aura period, active frequency band analysis in the EEG-Aura period, and mapping the reflections on the brain surface of the systematic changes occurring in the EEG-Aura period. Epilepsy is a chronic disorder that causes recurrent seizures due to hyper-synchronous discharges in the brain and is defined with neurobiological, cognitive, psychological, and social consequences as a result of these seizures. The generally uncertain nature of epilepsy and the often unpredictability of seizures hamper the quality of life of those exposed to this disease. Some patients may be considered fortunate compared to others in terms of predicting any seizures that will occur. These are patients with epileptic auras. Patients can predict some convulsive seizures by auras that is a retrospective feel. In this study, the reflections of potential aura phenomena in electroencephalography (EEG) signals, which is one of the main methods used in the diagnosis of epilepsy and pre-surgical evaluation, were examined and patient-specific EEG-Aura periods were determined. In addition to the determination of the EEG-Aura period, effective frequency band analysis and topographic mapping of the sub band activities were also performed. In determining the patient-specific EEG-Aura, one-dimensional EEG signals were represented in short-time Fourier transform (STFT) based spectrogram and continuous wavelet transform (CWT) based scalogram image form. Thus, EEG signals reflecting the amplitude values of the electrical activities of the brain over time are represented in the time-frequency domain in terms of energy density functions, and their spatial dimension has been increased. By applying the proposed ESA algorithm to the EEG signals represented in the image form, the systematic changes occurring before the epileptic seizure were defined and the subjective EEG-Aura period was determined for each patient. Also, spectrogram and scalogram based performances were compared in EEG-Aura detection. Considering all patients evaluated, consistency was observed between the patient-specific EEG-Aura periods determined for both methods. Considering the results obtained for all patients, it was determined that the 1-min interval covering the time from the second minute to the third minute before the seizure had the highest mean and the lowest variance to determine the systematic changes before the seizure. The overall success for the EEGAura period was evaluated in terms of accuracy, specificity, sensitivity, precision, and f-score model performance criteria. The average accuracy values in EEG-Aura detection with spectrogram and scalogram based ESA model was obtained as 90.81 ± 5.60% and 90.55 ± 5.40%, respectively. In the frequency band analysis in the EEG-Aura period, it was tried to determine the powerful frequency range in the classification stage by evaluating different frequency bands separately. Also, in determining the effective frequency band, the findings obtained regarding spectrogram and scalogram images were compared. For the EEG-Aura period, it has been observed that 0.5-30Hz and gamma-band activities can exhibit a successful performance and these frequency dynamics evaluated as effective frequency ranges. For 0.5-30Hz and gamma band, spectrogram and scalogram based average classification accuracies were 94.09 ± 5.77%, 97.42% ± 2.48 and 97.26 ± 3.39%, 96.62 ± 3.57%, respectively. The reflections of the neurophysiological changes occurring in the EEG-Aura period on the brain surface were mapped topographically and the brain regions where the changes occurred were determined. When the findings obtained for all patients were examined, it was observed that their low-frequency activities were generally effective in the frontal lobe, while high-frequency activities were effective in the temporal lobe. However, to examine the relationship between epileptic aura and EEG-Aura, reflections have to be handled patient-specific. It is considered that the findings obtained by comprehensively evaluating the novel approach-based preictal activities proposed in this thesis will be an important reference, especially for seizure prediction studies. Also, with the proposed model, it is predicted that systems can be developed in which epilepsy patients who do not respond to medical treatment methods will integrate more comfortably into life.